AI got good at code first because code comes with a cheap grader. The tests pass or they fail, and reinforcement learning runs on that signal. Fiction never had a cheap one, and I suspect that's a big reason model-written stories so often sound alike. Pocket FM thinks it found one. It publishes the stories and runs the app people hear them in. So it sees minute by minute where listeners drop off and whether they come back for the next episode. Its AI writing tool, Sherpa, went into free beta…